Term-structure models using binomial trees 9780943205533, 0943205530

Citation preview

Gerald W. Buetow, Jr., CFA BFRC Services, LLC James Sochacki James Madison University

C

H

FO UN

TM

D

N

R ES E A

IO AT

R

Term-Structure Models Using Binomial Trees

O

F AIMR

The Research Foundation of AIMR™

Research Foundation Publications Active Currency Management by Murali Ramaswami Common Determinants of Liquidity and Trading by Tarun Chordia, Richard Roll, and Avanidhar Subrahmanyam Company Performance and Measures of Value Added by Pamela P. Peterson, CFA, and David R. Peterson Controlling Misfit Risk in Multiple-Manager Investment Programs by Jeffery V. Bailey, CFA, and David E. Tierney Corporate Governance and Firm Performance by Jonathan M. Karpoff, M. Wayne Marr, Jr., and Morris G. Danielson Country Risk in Global Financial Management by Claude B. Erb, CFA, Campbell R. Harvey, and Tadas E. Viskanta Country, Sector, and Company Factors in Global Equity Portfolios by Peter J.B. Hopkins and C. Hayes Miller, CFA Currency Management: Concepts and Practices by Roger G. Clarke and Mark P. Kritzman, CFA Earnings: Measurement, Disclosure, and the Impact on Equity Valuation by D. Eric Hirst and Patrick E. Hopkins Economic Foundations of Capital Market Returns by Brian D. Singer, CFA, and Kevin Terhaar, CFA Emerging Stock Markets: Risk, Return, and Performance by Christopher B. Barry, John W. Peavy III, CFA, and Mauricio Rodriguez The Founders of Modern Finance: Their PrizeWinning Concepts and 1990 Nobel Lectures Franchise Value and the Price/Earnings Ratio by Martin L. Leibowitz and Stanley Kogelman

Global Asset Management and Performance Attribution by Denis S. Karnosky and Brian D. Singer, CFA Interest Rate and Currency Swaps: A Tutorial by Keith C. Brown, CFA, and Donald J. Smith Interest Rate Modeling and the Risk Premiums in Interest Rate Swaps by Robert Brooks, CFA The International Equity Commitment by Stephen A. Gorman, CFA Investment Styles, Market Anomalies, and Global Stock Selection by Richard O. Michaud Long-Range Forecasting by William S. Gray, CFA Managed Futures and Their Role in Investment Portfolios by Don M. Chance, CFA The Modern Role of Bond Covenants by Ileen B. Malitz Options and Futures: A Tutorial by Roger G. Clarke Risk Management, Derivatives, and Financial Analysis under SFAS No. 133 by Gary L. Gastineau, Donald J. Smith, and Rebecca Todd, CFA The Role of Monetary Policy in Investment Management by Gerald R. Jensen, Robert R. Johnson, CFA, and Jeffrey M. Mercer Sales-Driven Franchise Value by Martin L. Leibowitz Time Diversification Revisited by William Reichenstein, CFA, and Dovalee Dorsett The Welfare Effects of Soft Dollar Brokerage: Law and Ecomonics by Stephen M. Horan, CFA, and D. Bruce Johnsen

Term-Structure Models Using Binomial Trees

To obtain the AIMR Product Catalog, contact: AIMR, P.O. Box 3668, Charlottesville, Virginia 22903, U.S.A. Phone 434-951-5499; Fax 434-951-5262; E-mail [email protected] or visit AIMR’s World Wide Web site at www.aimr.org to view the AIMR publications list.

The Research Foundation of The Association for Investment Management and Research™, the Research Foundation of AIMR™, and the Research Foundation logo are trademarks owned by the Research Foundation of the Association for Investment Management and Research. CFA®, Chartered Financial Analyst™, AIMR-PPS ®, and GIPS ® are just a few of the trademarks owned by the Association for Investment Management and Research. To view a list of the Association for Investment Management and Research’s trademarks and a Guide for the Use of AIMR’s Marks, please visit our Web site at www.aimr.org. © 2001 The Research Foundation of the Association for Investment Management and Research All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior written permission of the copyright holder. This publication is designed to provide accurate and authoritative information in regard to the subject matter covered. It is sold with the understanding that the publisher is not engaged in rendering legal, accounting, or other professional service. If legal advice or other expert assistance is required, the services of a competent professional should be sought. ISBN 0-943205-53-0 Printed in the United States of America November 2001

Editorial Staff Roger S. Mitchell Book Editor Lisa S. Medders Assistant Editor

Jaynee M. Dudley Production Manager

Donna C. Hancock Production Coordinator

Kelly T. Bruton/Lois A. Carrier Composition

Mission

The Research Foundation’s mission is to

identify, fund, and publish research that is relevant to the AIMR Global Body of Knowledge and useful for AIMR member investment practitioners and investors.

Biographies Gerald W. Buetow, Jr., CFA, is currently president of BFRC Services, LLC, a quantitative financial consulting firm in Charlottesville, Virginia. Previously, he was the Wheat First Professor of Finance and director of the quantitative finance program at James Madison University. He was also lead quantitative researcher for Prudential Investment's Quantitative Investment Management Group, where he managed enhanced index funds and developed structured securities. Mr. Buetow has recently worked on developing financial modeling software for risk management and contingent claim valuation. He has published numerous articles in academic and practitioner journals as well as in various edited works. Mr. Buetow also recently completed a book with Frank J. Fabozzi titled Valuation of Interest Rate Swaps and Swaptions. He holds a B.S. in electrical engineering, an M.S. in finance from the University of Texas at Dallas, and a Ph.D. in finance and econometrics from Lehigh University and has extensive training in nuclear engineering. Mr. Buetow can be contacted via his Web site at www.BFRCServices.com. James Sochacki is an associate professor and director of the Center for Computational Mathematics and Modeling in the mathematics department at James Madison University. He has spent several years working with the quantitative finance program at James Madison University and taught mathematical finance several times. Mr. Sochacki developed many of the applied and computational mathematics courses at James Madison University. He has published articles on scientific and financial computation and modeling. Mr. Sochacki has also directed and consulted on many scientific and financial projects in academia, industry, and business. He holds a Ph.D. in applied mathematics from the University of Wyoming.

vi

©2001, The Research Foundation of AIMR™

Contents Foreword . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

viii

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

x

Chapter 1. Chapter 2. Chapter 3. Chapter 4. Chapter 5. Appendix A. Appendix B. Appendix C.

Interest Rate Modeling. . . . . . . . . . . . . . . . . . . . . . . Models for the Short Rate . . . . . . . . . . . . . . . . . . . . The Binomial Trees . . . . . . . . . . . . . . . . . . . . . . . . . Examples of the Binomial Models . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Variable Time Step . . . . . . . . . . . . . . . . . . . . . . . . . . Continuous Compounding of Interest Rates . . . . . Implementation and Calibration Issues for the Binomial Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1 11 34 57 80 81 83

Selected AIMR Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

94

84 91

Foreword Term-structure models are essential for the valuation of interest rate dependent claims. Although term-structure experts have produced a variety of useful models, they involve complex mathematics, which limits their accessibility to investment practitioners who are not engaged in this area of specialization. Moreover, the original “journal” versions of these models and their subsequent descriptions in text books often abstract from many important details necessary for implementation. These circumstances make it difficult for investors to compare the prices of interest rate dependent claims, to assess the appropriateness of alternative term-structure software products, and to build their own term-structure models. With this monograph, Gerald W. Buetow, Jr., CFA, and James Sochacki go a long way toward ameliorating this problem. They begin with a concise but hardly superficial overview of interest rate modeling, and they introduce the binomial tree framework. Having thoroughly prepared the reader, they next present the five most important no-arbitrage term-structure models: •

Ho–Lee Model. This model was the first no-arbitrage term-structure model. It assumes constant and identical volatility for all spot and forward rates and does not incorporate mean reversion.

•

Hull–White Model. This model extends the Ho–Lee model to allow for mean reversion.

•

Kalotay–Williams–Fabozzi Model. This model assumes a lognormal distribution and eliminates the problem of negative short rates, which can occur with the Ho–Lee and Hull–White models.

•

Black–Karasinski Model. An extension of the Kalotay–Williams– Fabozzi Model, this model controls the growth in the short rate.

•

Black–Derman–Toy Model. This model permits independent and timevarying spot-rate volatilities.

Buetow and Sochacki explain clearly and precisely the strengths and weaknesses of each model and how these models relate to one another. In addition, they generate binomial trees for each of the models to demonstrate the implications of their different features. Perhaps of most interest to the investment practitioner, they provide all of the detailed information necessary viii

©2001, The Research Foundation of AIMR™

Foreword

to implement these models, along with an easy-to-follow numerical example. For the first time in a single source, Buetow and Sochacki remove the opaqueness of the important no-arbitrage term-structure models, rendering them available for use by investment practitioners. The Research Foundation is pleased to present Term-Structure Models Using Binomial Trees. Mark Kritzman, CFA Research Director The Research Foundation of the Association for Investment Management and Research

©2001, The Research Foundation of AIMR™

ix

Preface This monograph was a by-product of the Fixed-Income Specialization Project (FISP) sponsored by AIMR. Throughout the FISP and the corresponding application development, called the Fixed-Income Specialization Application Project (FISAP), it became evident that most users of interest rate models do not have a very good understanding of them. This is largely due to the lack of available literature in this area. More specifically, existing readings require the reader to have an advanced understanding of mathematics. The demand for a readable source was evident from our discussions with both the participants and other investment professionals. This monograph is our attempt to meet that demand. While mathematics is unavoidable in this area, we have tried to concentrate more on the finance so that the reader will not get lost in the algebra. We have also tried to outline the specifics required if an analyst needs to implement his or her own model. We hope this monograph provides an intelligible and useful guide on term-structure models for readers, thus eliminating the frustration often associated with this area.

In Memory of

Captain William F. Burke, Jr. Engine Company 21, FDNY A Leader, Comedian, Friend, and Hero You paid the supreme sacrifice to ensure the safety of your men and countless others on the morning of September 11, 2001. Men of your courage, leadership, kindness, intelligence, and humor are far too rare in this world. I am forever indebted to you for helping me navigate the labyrinth of adolescence and helping me solve the numerous personal and professional dilemmas that come with age. The current value of that counsel has no metric. My memories of our friendship are more precious than a pure arbitrage. You will be missed but never forgotten. GWB x

©2001, The Research Foundation of AIMR™

1. Interest Rate Modeling Many models are available for modeling the short rate—that is, the forward rate for the short time period—and many papers and books have been written on these models. Models for the short rate come in various forms: Noarbitrage, equilibrium, Markov, and non-Markov are some examples. These frameworks often involve intimidating mathematics. For the practitioner, the no-arbitrage models seem to be the most widely accepted, but unfortunately, implementing these models can be challenging. This monograph explains many of the financial and mathematical features of no-arbitrage models. Using any of these models on security or contingent-claim valuation is not as straightforward as it first may appear. What makes implementation even more difficult is the lack of guidance in the extant literature. The original articles assume, when it comes to actual implementation, that the user has a lot of mathematical knowledge of the models. That is, when a user actually sits down to generate the appropriate code to implement the model, the articles are found wanting. Jamishidian (1991), Rebonato (1996), Tuckman (1996), Bjerksund and Stensland (1996), Sundaresan (1997), and Hull (2000) offer some additional insight into implementation. Still, these studies do not explain either the equations or the properties of the models in enough detail to give the user the information needed to implement the models properly. We hope to fill that need with this work. In this monograph, we develop the models and the mathematical equations and approximations for them, and we give examples of the models. We then present the strengths and weaknesses of the models and indicate how the models can be used to model the short rate. First, we develop the terminology and notation used in this monograph. Because interest rates are stated at certain times and act over a period of time, we need to formulate how time is measured. Calendar time is measured in years from a certain starting time. That is, if we start measuring interest rates on January 15, 1970, then t = 0 on January 15, 1970, t = 1 on January 15, 1971, t = 2 on January 15, 1972, and so on. For example, the interest rate may be effective at t = 1 for a time period of six months. We can divide the calendar time into increments called “time steps.” If the time step is six months, or 0.5 years, then there are two time steps in t = 1 and four time steps in t = 2. We number the time steps consecutively. The time steps do not have to be of the same length. For example, the first time step could be six months, the second

©2001, The Research Foundation of AIMR™

1

Term-Structure Models Using Binomial Trees

could be five months, and the third could be four months. In that case, t = 1 is in the third time step. We always choose t0 = 0 as the starting time for our interest rate models. If the calendar times for modeling the interest rate are chosen to be 0 = t0 < t1 < t2 A2 ( 1

+ R1τ )

< A2(1

+ R 1 τ ),

or P2

---------------------

(1

+ r1 τ )

then an arbitrage opportunity exists. That is, a risk-free profit is available without any initial investment. Because the models we are considering are noarbitrage models, Equation 6 must be satisfied. In particular, from Equation 6, we must have 1

--------------------------

(1

+ R2 τ )

2

=

=

1 ⁄ ( 1 + r1 τ )

----------------------------

1 + R1 τ

(7)

1 ⁄ ( 1 + r1 τ )

---------------------------- .

1 + r0 τ

Because the short rate is the forward rate at a given t for a short time period, we can denote it as r(t). That is, r(t ) is the short (forward) rate at time t for a short time period, τ. This short time period for the short rate is usually fixed over calendar time. In the models we consider, the short rates are usually for a fixed six-month period. Financial analysts are interested in determining possible short rates at future points in time. For example, they are interested in knowing what the six-month short rate will be seven years from now. Suppose the graph in Figure 1 is a plot of the six-month (fixed-time-period) short rate for a calendar 4

©2001, The Research Foundation of AIMR™

Interest Rate Modeling

Figure 1.

Six-Month Short Rate (Forward Rate) for 30 Years

Six-Month Short Rate 0.12

0.10

0.08

0.06

0.04

0.02

0 0

2

4

6

8

10

12

14

16

18

20

22

24

26

28

30

Years

time of 30 years. The graph indicates that the six-month short rate at Year 7 is 6 percent, which we would write as r(7) = 0.06.2 The graph in Figure 1 shows that the short rate can have significant up and down movement. This movement is called “the volatility of the short rate.” Figure 1 also shows that the volatility is a function of the calendar time, t. Therefore, a model of the interest rate must be able to incorporate volatility. The models we present incorporate the volatility of the short rate, which we call “the local volatility” to distinguish it from yield volatility. The no-arbitrage models we discuss attempt to answer the following question: Can a model be developed that incorporates the parameters affecting the short rate that will produce graphs consistent with the actual short rate? An elementary knowledge of calculus and probability is all that is necessary to learn and understand the processes that are commonly used to model the short rate within the no-arbitrage framework. The model must make intuitive sense.

2 For

a semiannual model, when t = 7 years and τ = 0.5 years, there are 14 periods.

©2001, The Research Foundation of AIMR™

5

Term-Structure Models Using Binomial Trees

The models we present are some of the most commonly used models in the finance literature. Our objective is for readers to understand the models so that they can implement them properly. Once analysts understand the appropriate use of the models, they can apply them to contingent-claim valuation, as in Buetow and Fabozzi (2001), or to risk-metric computation, as in Buetow and Johnson (2000).

Introduction to Binomial Models Consider the following simple example of a volatile short rate for a two-period bond that pays $100 at the end of the second time period. Because many unpredictable events can change market conditions (including short rates and spot rates), the spot rates at time t = τ will be different from the spot rates at time t = 0. Suppose we believe that at time t = τ, the short rate for the next time period can have only one of two possible values, ru or rd, with ru > rd. If there is a probability q that the short rate will be ru, then there is a probability 1 – q that the short rate will be rd (because the total of the probabilities must be 1). With these conditions in mind, from Equation 2, the present value of the bond is A2 =

$100

--------------------------

(1

+ R2τ )

2

,

and from Equation 4, the present value at time period τ is either $100 1 + ruτ

pu =

-----------------

pd =

-----------------

or $100 . 1 + rd τ

Therefore, the present value at t = 0 is either pu

-----------------

1 + r0 τ

or pd

-----------------

1 + r0 τ

.

Thus, within the no-arbitrage framework, either 6

©2001, The Research Foundation of AIMR™

Interest Rate Modeling

A2 =

$100

--------------------------

(1

+ R2τ )

2

pu

(8a)

=

-----------------

=

--------------------------------------

1 + r0 τ

$100 ⁄ ( 1 + r u τ ) 1 + r0 τ

or A2 =

$100

--------------------------

(1

+ R2τ )

2

pd

(8b)

=

-----------------

=

-------------------------------------- .

1 + r0 τ

$100 ⁄ ( 1 + r d τ ) 1 + r0 τ

Equations 8a and 8b are similar to Equation 7. A2 is now set equal to the expected value of Equations 8a and 8b, which gives A2 =

$100

--------------------------

(1

+ R2τ )

2

p p = q ----------u------- + ( 1 – q ) ----------d-----1 + r0 τ 1 + r0 τ $100 ⁄ ( 1 + r τ ) $100 ⁄ ( 1 + r τ ) = q -------------------------------u------- + ( 1 – q ) -------------------------------d------1 + r0 τ 1 + r0 τ

or, after dividing by $100 (the value of the bond), 1

-------------------------- =

(1

+ R2 τ )

2

1 ⁄ (1 + r τ ) 1 ⁄ (1 + r τ ) q ----------------------u------- + ( 1 – q ) ----------------------d------ . 1 + r0τ 1 + r0τ

Clearing fractions leads to 1

-------------------------- ( 1

(1

+ R2 τ )

2

+ r0 τ ) ( 1 + rd τ ) ( 1 + ru τ ) = q ( 1 + rd τ ) + ( 1 – q ) ( 1 + ru τ ) .

Finally, subtracting the right-hand side from both sides produces

©2001, The Research Foundation of AIMR™

7

Term-Structure Models Using Binomial Trees

1

-------------------------- ( 1

(1

+ R2 τ )

2

+ r0 τ ) ( 1 + rd τ ) ( 1 + ru τ ) – q ( 1 + rd τ ) –( 1 – q ) ( 1 + ru τ ) = 0

(9a)

or A2(1 + r0τ)(1 + rdτ)(1 + ruτ ) – q(1 + rdτ) – (1 – q)(1 + ruτ) = 0.

(9b)

Equation 9b is the fundamental equation used in no-arbitrage binomial models to determine the possible short rates. To use Equation 9b, one has to model ru and rd. In the following chapters, we discuss five models that give formulas for ru and rd. The models are the Ho–Lee (HL; Ho and Lee 1986), Hull–White (HW; Hull and White 1994), Kalotay–Williams–Fabozzi (KWF;3 Kalotay, Williams, and Fabozzi 1993), Black–Karasinski (BK; Black and Karasinski 1991), and Black–Derman–Toy (BDT; Black, Derman, and Toy 1990). These models build on Equation 9b. Historically, the HL model was introduced first, then the BDT model, then the BK model, then the KWF model, and then the HW model. We, however, present the models in such a way that each one arises mathematically from the previous model. Thus, for example, a computer program designed to implement the HL model could easily be adapted for the other models. In Chapter 3, we explain binomial short-rate models and their corresponding price trees. We cover many of the commonly used models for several reasons. First, each model has different properties, and these properties must be understood before the model is applied for valuation purposes. Second, the models require different input parameters. These inputs must also be properly understood before the model is implemented. Third, all the models have limitations that need to be appreciated prior to their use. Finally, interpreting outputs (values or risk metrics) from the model is impossible unless the underlying model and its corresponding characteristics are understood.4 Our goal for this monograph is to help readers gain an appreciation for the complexity and power of binomial models. But keep in mind that, although binomial models require far less computing than higher-order tree models, such as trinomial and tetranomial models, these higher-order models do have some advantages, which the reader should also appreciate after completing this monograph. Fortunately, despite the interest rate lattice used in the higher-order models, the mechanics are similar to the mechanics of the 3 The

actual KWF model is a slight variation of the model we present. Other names are also associated with this model. 4 Many

8

other reasons exist, but they do not need to be listed here.

©2001, The Research Foundation of AIMR™

Interest Rate Modeling

binomial models, so the concepts developed here are easily extrapolated to the other lattice structures. What we develop in this monograph can also be straightforwardly modified to work for trinomial and tetranomial models. A trinomial model would have three possible values for the short rate r1—say, ru, rm, and rd—with the respective probabilities qu, qm, and qd (where qu + qm + qd = 1). Equation 9b would become A2(1 + r0τ)(1 + rdτ)(1 + rmτ)(1 + ruτ) – qu(1 + rdτ)(1 + rmτ) – qm(1 + rdτ)(1 + ruτ) – qd(1 + rmτ)(1 + ruτ) = 0.

(10)

We briefly discuss trinomial models at the end of Chapters 3 and 4. The models we present are the basic models. Advanced readers may suggest that the more sophisticated models are more useful. Undoubtedly, they are. But the more complicated models cannot be properly understood without an understanding of the fundamental models. Moreover, despite the refinements of the sophisticated models, the models we discuss are still widely used by practitioners.

Uses of Term-Structure Models At this point, the reader may think that binomial models for the short rate are severely limited, but they are not. Binomial models exhibit many of the desirable properties of the short rate. Additionally, they are commonly used as valuation models for both bonds and fixed-income contingent claims.5 The concepts presented in this monograph have far-reaching applications throughout the investment arena. Almost every interest rate product in existence can be analyzed by using the models presented, so to reach an accurate conclusion about an investment, investment professionals must understand these models. Securities that particularly lend themselves to analysis by way of lattice models are those with option characteristics, such as swap options, caps, floors, callable bonds, putable bonds, bonds with sinking-fund provisions, bond options, options on bond futures, capped floaters, and credit options. Even securities without embedded options, however, can be analyzed within this framework as part of credit-spread or scenario analysis. The applicability of the models within the fixed-income area is almost endless. Another important application of these models for the investment professional is the measurement of interest rate sensitivity. Specifically, effective duration, effective convexity, partial duration, key-rate duration, option5 Two good sources on how to use short-rate binomial models and their corresponding price trees are Fabozzi (2000) and Buetow and Fabozzi.

©2001, The Research Foundation of AIMR™

9

Term-Structure Models Using Binomial Trees

adjusted spread, slope duration, and curvature duration are easily evaluated with these models. Before the proper interpretation of the model’s output, however, the model generating the metrics must be understood because each model generates different values. We cannot overemphasize this point. An investment professional must be very comfortable with the characteristics of these models before interpreting any of the valuation or sensitivity results.

10

©2001, The Research Foundation of AIMR™

2. Models for the Short Rate Short-term interest rates change through time, so we model the short-term rate using a mathematical process. Specifically, we use differential equations or differential processes to capture the dynamics of the short-term interest rate. The model must also incorporate empirical and intuitive characteristics of interest rates, which the models presented in this chapter attempt to do. The models we consider have the general form df [r(t)] = {θ(t) + ρ (t)g[r(t)]}dt + σ[r(t), t]dz,

(11)

where f and g are suitably chosen functions, θ (which we will show is the drift of the short rate) is determined by the market, and ρ (the tendency to an equilibrium short rate) can be chosen by the user of the model or dictated by the market. The term σ is the local volatility of the short rate. In Equation 11, dz = ε dt , where ε ~ N(0,1) and N(0,1) is the normal distribution with a mean of 0 and a standard deviation of 1.6 This component is incorporated in the differential equation to model the randomness (volatility) of interest rates.7 Differential equations with this type of term are called stochastic differential equations (SDEs). The models based on Equation 11 are known as “onefactor models” because they model only the short rate (one factor). The right-hand side of Equation 11 is the sum of two components. The first component is contained in the braces and is multiplied by the term dt. This component is the expected or average change in rates over a short time period, dt. This component is where certain characteristics of interest rates, such as mean reversion (MR), are incorporated. Interest rates are expected to change by what this term suggests. The second component models the randomness (or risk) of interest rates. This component is a product of the terms σ[r(t),t] and dz. This component dictates the distribution characteristics of interest rates. Depending on the model, interest rates are either normally or lognormally distributed. As explained, on average, this component is zero, which is why the first component, the expected value, needs to be thoroughly understood. process. This means dz = ε dt and that the values of ∆z for any two different time intervals ∆t are independent. 6 The variable z follows a Wiener

7 Throughout the manuscript, we will take liberties with terminology, but this will not detract from the financial implications of the development.

©2001, The Research Foundation of AIMR™

11

Term-Structure Models Using Binomial Trees

We will not study the technical issues of SDEs in this monograph, but we will use Equation 11 to develop five one-factor models for the short rate. To study Equation 11 in full detail, one needs a good background in stochastic calculus.8 Because analytic (or closed-form solutions) to Equation 11 are not feasible even when stochastic calculus is used, numerical approximations to Equation 11 that incorporate the stochastic term σ[r(t),t ]dz are usually applied. Numerical calculations can be done only at discrete calendar times; therefore, calendar times at which the short rate will be calculated must be chosen. In this monograph, the chosen times are denoted as indicated in Chapter 1, and the times are usually chosen so that the time steps will match the time period for the spot rates of the term structure. This approach makes it easier to incorporate the no-arbitrage condition (bond prices from the model must match the bond prices from the term structure of spot rates). We later explain, however, that some models do not allow the time step to match the time periods for the spot rates. In those cases, the analyst has to use interpolation to match the model time steps with the time periods of the term structure. We let r(tk ) denote the exact solution to Equation 11 at time t = tk. Because we cannot (in general) generate an analytic form for r(tk ), we develop a numerical (or discrete) process that will approximate r(tk ). We denote this approximation by rk. That is, our numerical algorithm will generate a value rk that approximates r(tk ). We write rk ≈ r(tk ). We now use calculus to develop a numerical approximation for Equation 11. We have that dr ≈ ∆r = r(tk+ 1) – r(tk ) ≈ rk+1 – rk (12a) when dt ≈ ∆t = tk+ 1 – tk

(12b)

is close to zero. In many cases, ∆r is a fairly good approximation of dr, even if dt is different from zero. (We will use rk both as r(tk ) and as the approximation of r(tk ). The context will make clear whether we are using the exact value or the approximation.) We will use τ for ∆t. If the time steps vary, τ will also depend on k. We use Equations 12 on Equation 11 to develop our numerical approximation for Equation 11. 8 Determining

solutions to Equation 11 is generally difficult. Stochastic calculus must be used to obtain closed-form solutions. A good textbook on the mathematical theory behind Equation 11 is Neftci (1996). For example, to focus attention on the first term of the SDEs (Equation 11), we state that we are taking the expectation of the SDE. We are not really applying the expectations operator to both sides of the equation. We are simply eliminating the second term of the SDE to analyze the first component of the equation. If we did apply the expectations operator, the only difference to our analysis would be a scaling factor; the dynamics do not change. Applying the operator unnecessarily complicates the analysis.

12

©2001, The Research Foundation of AIMR™

Models for the Short Rate

As stated earlier, because the expected value of the random term is zero, we analyze the properties of Equation 11 with only the first component included. In other words, we set the second component equal to its expected value of zero and study the resulting equation. Most of the properties of this resulting equation (without the stochastic term) will apply to Equation 11 (with the stochastic term). A basic understanding of calculus is all that is needed to study this equation. Inputs into the model must satisfy the mathematical properties of both the SDEs and their approximations. We assume the time step is constant and make adjustments when it is not. The five models are now developed.

Ho–Lee Model In the HL model or process f(r) = r, g(r) = 0, and ρ = 0 in Equation 11. The HL process is, therefore, given by dr = θdt + σdz.

(13)

Because dz = ε dt is normally distributed, the HL process models the short rate as normal. The solution to Equation 13, assuming r(0) = r0 is given by r ( t ) = r 0 + ∫ t θ ds + ∫ t σ dz , 0 0

(14a)

where the integral involving σ is a stochastic integral. If θ is constant, the solution can be expressed as r ( t ) = r 0 + θ t + ∫ t σ dz . 0

(14b)

Equation 14b shows that the HL process models an interest rate that can change proportionally with time t through the constant of proportionality θ and a random disturbance determined by σ. That is, the larger the magnitude of θ, the larger the average change in the short rate over time. This effect is why θ is called the drift in the short rate. Also, the smaller θ is, the larger the influence of the random disturbance is. A shortcoming of the HL process is that the short rate can be negative. Hull shows that θ is related to the slope of the term structure. A numerical approximation for Equation 13 can be obtained by using Equations 12a and 12b. Letting tk = kτ and rk ≈ r(kτ) gives rk+1 – rk = θkτ + σk ∆zk

as an approximation for Equation 13 or rk+1 = rk + θkτ + σk ∆zk,

(15)

where ∆zk is a numerical (discrete) approximation of dz. Because dz = ε dt , we can further approximate Equation 15 by ©2001, The Research Foundation of AIMR™

13

Term-Structure Models Using Binomial Trees

r k +1 = r k + θ k τ + σ k ε k τ ,

(16)

where εk is a random number generated from a normal distribution N(0,1). Equation 16 is the form of the expression that is used for rk+1 to build the HL binomial tree. In the HL process, Equation 16 is used for the formula for ru and rd in Equation 9. We address the issues of this approximation for the binomial tree in the next chapter. We first consider the expected value solution to Equation 16 when θ is constant. Equation 16 under these requirements is rk+1 = rk + τ θ,

(17a)

and the solution is given by r k = c + kδ ,

(17b)

where c and δ are constants. In particular, c = r0 = R1 and δ = θτ. The last equation shows that the mean short rate in the HL process increases or decreases at a constant rate θ over time, depending on the sign of θ. As a matter of fact, Equation 17b shows that the short rate grows without bound if θ > 0 and decreases without bound (i.e., becomes extremely negative) if θ < 0. The binomial tree examples considered in Chapter 4 illustrate this effect. For now, look at solutions to Equation 13 and Equation 16. We use a random number generator to model the normal distribution. Figure 2 depicts a Ho–Lee example with σ = 10 percent averaged over 1,000 simulations (the multiple runs line) and the corresponding expected value solution from Equations 16 and 17b with c = r0 = R1 = 0.05. The slope of the curve for the expected value solution is the drift, θ. As expected, the expected value solution curve and the curve for the average over 1,000 simulations have the same general trend. Clearly, the single random run is highly volatile. For the numerical approximations using Equation 16, we used τ = 0.5. (Unless specified otherwise, we use τ = 0.5 in all our numerical approximations.)

Hull–White Model In the HW model or process, f(r) = r, g(r) = r, and ρ = –φ. Therefore, the stochastic process for the HW model for the short rate is dr = (θ – φ r)dt + σdz.

(18)

The short-rate process in the HW models is seen to be normal as in the HL process. We consider the case where the parameters θ and φ are constant over time. Note that if φ = 0, the HW process reduces to the HL process. (The HW process will be similar to the HL process, if φ is close to zero.) The introduction 14

©2001, The Research Foundation of AIMR™

Models for the Short Rate

Figure 2.

Example Ho–Lee Solutions

Rate (%) 14 12 10 8 6 4 2 0 2 4 6 0

0.5

1.5

2.5

3.5

4.5

5.5

6.5

7.5

8.5

9.5

Years Expected Value Solution

Multiple Runs

Single Run

of φ in the HW model is an attempt to incorporate mean reversion and to correct for the uncontrolled growth (or decline) in the HL model; φ is the mean reversion rate. Note that, despite this term, the HW model can still result in negative interest rates. The expected value of Equation 18 leads to the ordinary differential equation dr = (θ – φ r)dt,

(19)

whose solution is given by r(t) =

–φ t θ -- + ce , φ

(20)

where θ c = r 0 – -- . φ ©2001, The Research Foundation of AIMR™

(21)

15

Term-Structure Models Using Binomial Trees

If φ > 0, we see from Equation 20 that -lim r ( t ) = θ φ

t →∞

= µ.

Therefore, for positive mean reversion ( φ > 0), the HW process will converge to the short rate, µ = θ/φ. Thus, the term µ is called the target or long-term mean rate. For negative mean reversion (φ < 0) the short rate grows exponentially over time. Equation 21 indicates that c in the HW process gives the distance of the initial short rate from the target short rate. Therefore, the closer the initial short rate is to the target rate, the smaller c will be and the smaller changes in r(t) will be through time. In particular, Equation 21 shows that if r(0) = µ, then c = 0, and from Equation 20 that r(t) = µ for all t. That is, under this scenario, the short rate is always the target rate. Consider the following examples that give qualitative information for c and µ. Factoring φ in Equation 18 leads to dr = φ (µ – r)dt + σdz,

the expected value of which leads to dr = φ (µ – r)dt.

We see that if r > µ, then dr is negative and r will decrease, and if r < µ, then dr is positive and r will increase. That is, r will approach the target rate µ. The larger φ is, the faster this approach is to the target rate µ. This is why φ is called the mean reversion or mean reversion rate. It regulates how fast the target rate is reached. Since the target rate µ is equal to θ/φ, we can solve for the drift, θ, or the mean reversion, φ. That is, θ = µφ

(22)

or φ =

θ . µ ---

(23)

It is seen from Equation 22 and Equation 23 that there is a strong relationship between the drift and mean reversion that can be used to reach any desired target rate. How large the mean reversion should be is an important financial question. Equations 22 and 23 can be used to set target

16

©2001, The Research Foundation of AIMR™

Models for the Short Rate

rates. Equations 20 and 21 allow one to determine how long it takes to reach the target rate. We consider examples of these properties below. Using Equations 12a and 12b to approximate Equation 18 gives us r k + 1 = r k + ( θ k – φ k r k )τ + σ k ε k τ .

(24)

If θ and φ are constant, then the solution to the expected value of Equation 24 has the form rk = αβk + γ.

To determine α, β, and γ, we substitute this form for rk into Equation 24 without the stochastic component and obtain that β = (1 – φτ), γ = θ/φ = µ, and α = r0 – µ. Therefore, k θ r k = α ( 1 – φτ ) + -- . φ

(25)

Note that if 0 < φτ < 2, then –1 < 1 – φτ < 1 and θ lim r k = -φ

k →∞

= µ,

which is the same result that is obtained from Equation 20 for the HW SDE under the same scenario. As will be seen later, the condition 0 < φτ < 2 can be maintained in modeling the short rate. Equation 24 is the formula that will be used for ru and rd in Equation 9 for the HW process. Figure 3 illustrates the results for five HW examples, with r(0) = 5 percent; φ = 0.02 (denoted as “mean reversion” in Figure 3 and other figures); µ = 6.5 percent; and σ = 5 percent, 10 percent, 15 percent, and 20 percent averaged over 1,000 simulations. Figure 3 also includes the expected value solution (σ = 0) from Equations 20 and 21, with c = –0.015. The drift from Equation 22 is given by θ = µφ = (0.065)(0.02) = 0.0013. Again, as expected, the average over 1,000 simulations is close to the expected value solution. Note that the graphs have not reached the target rate of 6.5 percent, because a mean reversion of 0.02 is small. Figure 4 presents five HW examples with r(0) = 0.05; φ = 0, 0.01, 0.02, 0.05, and 0.15294; and σ = 10 percent. The results presented are an average of 1,000 simulations. Because r(0) = 0.05 and µ = 0.065 in all the examples, c is the same for each case and is equal to –0.015 (from Equation 21). The values for θ for each case can be calculated from Equation 22. From Equations 20 and 21, φ can be used to determine the time of convergence to the target rate.

©2001, The Research Foundation of AIMR™

17

Term-Structure Models Using Binomial Trees

Figure 3.

Example Hull–White Solutions (mean reversion equals 0.02; volatility varies)

Rate (%) 7

6

5

4 0

0.5

1.5

2.5

3.5

4.5

5.5

6.5

7.5

8.5

9.5

Years Expected Value

Volatility = 15%

Volatility = 5%

Volatility = 20%

Volatility = 10%

We use Equations 20 and 21 to calculate the time it takes to get within 5 percent of the target rate for each value of φ, as shown in Table 1. Choosing a meaningful value for φ is critical in modeling interest rates. For example, in Figure 4, mean reversion of 0.15294 is shown because this value for φ is necessary to get within 5 percent of the target rate after 10 years. Figure 4 shows this effect, and the curve with φ = 0 grows as expected. The value of φ clearly has a strong impact on the convergence to the target rate. As pointed out earlier, if the initial rate and target rate were closer, the convergence time would also drop. Table 1 shows the drift values and convergence times when the target rate is 5.75 percent. Table 1 clearly demonstrates that the relationship between r(0) and µ significantly affects the convergence time.

Kalotay–Williams–Fabozzi Model For the KWF process, f(r) = ln(r) (where ln is the natural logarithm with base e ≈ 2.718), g(r) = 0, and ρ = 0 in Equation 11. Thus, the differential process is dln(r) = θdt + σdz.

(26a)

This model is directly analogous to the HL model. Setting u = ln(r) in equation 26a, the HL process (Equation 13) for u is du = θdt + σdz.

18

(26b)

©2001, The Research Foundation of AIMR™

Models for the Short Rate

Figure 4.

Example Hull–White Solutions (mean reversion varies; volatility equals 10 percent)

Rate (%) 7.0

6.5

6.0

5.5

5.0

4.5 0

0.5

1.5

2.5

3.5

4.5

5.5

6.5

7.5

8.5

9.5

Years MR = 0

MR = 0.05

MR = 0.01

MR = 0.15294

MR = 0.02

Table 1.

Determining the Convergence Time to Target Rate for the Hull–White Model Target Rate (µ) = 6.5%

Mean Reversion (φ)

Drift Rate (θ)

Time to Converge to Target Rate (years)

0.01

0.00065

152.94

0.02

0.0013

76.47

Target Rate (µ) = 5.75% Drift Rate (θ)

Time to Converge to Target Rate (years)

0.000575

95.86

0.0015

47.94

0.05

0.00325

30.59

0.002875

19.18

0.15294

0.00994

10.0

0.00879

6.27

©2001, The Research Foundation of AIMR™

19

Term-Structure Models Using Binomial Trees

Because u follows a normal process, ln(r) follows a normal process, and so r follows a lognormal process. Because u is the same as the HL and HW processes, u can become negative, but u = ln(r) so r = eu; thus, r is always positive. Therefore, the KWF model eliminates the problems of negative short rates that occur with the HL and HW models. Taking the expectation of Equation 26a results in dln(r) = θdt,

and for Equation 26b du = θdt.

The expected value in Equation 14a yields ln [ r ( t ) ] = u = u ( 0 ) + ∫ t θ ds, 0

and because u(0) = ln[r(0)] = ln(r0), ln [ r ( t ) ] = ln ( r 0 ) + ∫ t θ ds . 0

Taking the exponential of both sides of this equation gives t

r ( t ) = r0 e

∫0

θ ds

,

(27)

showing that r(t) > 0 because r(0) > 0. Therefore, for constant θ, if θ > 0, the short rate in the KWF process grows without bound, and if θ < 0, the short rate in the KWF process decays to 0. From Equation 16, the discrete approximation of Equation 26b is uk + 1 = uk + θk τ + σkε k τ ,

(28a)

and the exponential of this equation gives the discrete approximation of Equation 26a: rk + 1 = rke

θk τ + σk ε k τ

.

(28b)

From Equation 28b, we see that the numerical approximation of the expected value for Equation 26a has similar properties to the solution to the KWF SDE. That is, for constant θ, if θ > 0, the short rate grows without bound, and if θ < 0, the short rate decays to zero. Equation 28b is the formula we will use for ru and rd in Equation 9 for the KWF process. 20

©2001, The Research Foundation of AIMR™

Models for the Short Rate

Figure 5 is similar to Figure 2 and presents a KWF model example with

σ = 10 percent for one random run, σ = 10 percent averaged over 1,000 random

simulations, and the expected value solution from Equations 27, with c = lnr(0) = ln(0.05) = –2.995. The value we use for the drift is –0.1366684. Note that the KWF random run follows the expected value solution more closely than the HL example did in Figure 2. This pattern is also exhibited in the binomial tree models in Chapter 3. The actual KWF model is a variation of the model we have presented. Because the actual KWF model does not explicitly incorporate the drift, the model does not always have a solution for the binomial tree. Our presentation of the KWF model, however, is easily modified to obtain the actual KWF binomial tree.

Figure 5.

Example Kalotay–Williams–Fabozzi Solutions

Rate (%) 6

5

4

3

2

1

0 0

0.5

1.5

2.5

3.5

4.5

5.5

6.5

7.5

8.5

9.5

Years Expected Value Solution

©2001, The Research Foundation of AIMR™

Single Run

1,000 Simulations

21

Term-Structure Models Using Binomial Trees

Black–Karasinski Model In the BK model, we set f (r) = ln(r), ρ = – φ, and g(r) = ln(r) in Equation 11 to obtain the SDE dln(r) = [θ – φ ln(r)]dt + σdz.

(29a)

We now work with Equation 29a using Equation 18 for the HW process in a manner similar to the way we used results from the HL process to develop the KWF process. Setting u = ln(r) in Equation 29a, we obtain du = (θ – φ u)dt + σ dz,

(29b)

which is the HW process for u. Again, note that u has all the same properties as r in the HW model. Because r = eu in the BK process, r > 0 in the BK process. This difference is the advantage the BK model has over the HW model. Thus, the BK process is an extension of the KWF process, just as the HW process is an extension of the HL process. The main difference is that the BK process is a lognormal extension of the lognormal KWF process. As a matter of fact, if φ= 0, the BK process reduces to the KWF process. The introduction of φ by Black and Karasinski controls the growth of the short rate in the KWF process. From Equation 20, we have u(t) =

–φ t θ -- + ce , φ

and after taking exponentials, r(t) = e = e

u (t ) ( θ ⁄ φ ) + ce

–φ t

(30) .

For φ < 0, we see that r grows without bound and that for φ > 0, lim r ( t ) = e

θ⁄φ

t →∞

= µ.

The target rate for the BK process is the exponential of the target rate for the HW process. As in the HW process from Equation 30 (or Equations 20 and 21), we see that θ c = ln ( r 0 ) – -φ

22

(31)

©2001, The Research Foundation of AIMR™

Models for the Short Rate

in the BK process. Again, the closer the initial rate is to the target rate, the faster the BK process converges to the target rate. From Equations 30 and 31, we see that if the initial short rate is the target rate, then r(t) = µ for all t in the BK process, which is analogous to the result we found for the HW process. As we did for the HW SDE, we present some examples that give qualitative information about c and µ for the BK model. Given the target rate µ = e θ ⁄ φ , we can solve for the drift or the mean reversion similarly to Equations 22 and 23 in the HW model. As before, we have θ = φ ln(µ)

(32)

and φ =

θ . ln ( µ )

(33)

-------------

Later, Equations 32 and 33 will be used to set target rates in the BK models so that we can compare the target rates of Figure 4 with the HW model. Equations 30 and 31 are used to determine how long it takes to reach the target rate. We discretize u = ln(r) in Equation 29b, just as we did for the HW SDEs, and then let r = eu. This approach is analogous to the way we used the HL discrete process to get the KWF discrete process. The equations corresponding to Equation 24 are u k + 1 = u k + ( θ k – φ k u k )τ + σ k ε k τ

(34a)

or, after taking the exponential of both sides of Equation 34a, r k +1 = r k e

[ θ k – φ k ln ( rk ) ] τ + σ k εk τ

.

(34b)

Similar to Equation 25, if θ and φ are constant, the solution to the expected value in Equation 34b is k

rk = e

α ( 1 – φ τ) + ( θ ⁄ φ )

.

(35)

Note that from Equation 35, lim r k = e k →∞

θ⁄φ

= µ

©2001, The Research Foundation of AIMR™

23

Term-Structure Models Using Binomial Trees

for 0 < φτ < 2. This result is similar to what we obtain from Equation 25 for the HW SDEs. Equation 34b is the formula we use for ru and rd in Equation 9 for the BK process within the binomial framework. Figure 6 corresponds to Figure 3 and presents five BK solutions with r(0) = 0.05; φ = 0.02; µ = 6.5 percent; σ = 5 percent, 10 percent, 15 percent, and 20 percent averaged over 1,000 simulations; and the expected value solution from Equations 30 and 34 with c = ln(r0) – ln(µ) = –0.26236. From Equation 32, the drift is –0.05467. Note that after 10 years, the target rate of 6.5 percent has not been reached. Figure 7 presents five BK examples with r(0) = 0.05; φ =0, 0.01, 0.02, 0.05, and 0.1632; and σ = 10 percent. Figure 7 for the BK SDE is analogous (under the u substitution) to Figure 4 for the HW SDEs. The results presented are an average of 1,000 simulations. The drifts are chosen from Equation 32 to match the target rate of 6.5 percent. Note that only the graph with φ = 0.1632 is approaching the target rate after 10 years because 0.1632 is the only value for φ that has the BK SDE short rate get within 5 percent of the target rate Figure 6.

Example Black–Karasinski Solutions (mean reversion equals 0.02; volatility varies)

Rate (%) 5.6

5.5 5.4

5.3 5.2 5.1 5.0 0

0.5

1.5

2.5

3.5

4.5 Years

5.5

6.5

7.5

Expected Value

Volatility = 15%

Volatility = 5%

Volatility = 20%

8.5

9.5

Volatility = 10%

24

©2001, The Research Foundation of AIMR™

Models for the Short Rate

Figure 7.

Example Black–Karasinski Solutions (mean reversion varies; volatility equals 10 percent)

Rate (%) 7

6

5 0

0.5

1.5

2.5

3.5

4.5

5.5

6.5

7.5

8.5

9.5

Years MR = 0

MR = 0.05

MR = 0.01

MR = 0.1632

MR = 0.02

after 10 years. The value is determined from Equations 30 and 31. As in the HW examples, a target rate closer to 5 percent would greatly reduce the convergence time to the target rate for the mean reversion values. As in the HW SDE, it is crucial to choose the mean reversion so that convergence to the target rate is achieved at the desired time. Note that the BK models seem to be more stable than the HW models. We will compare our binomial models with these results in Chapter 3. These give us a base for comparison in the sections that follow.

Black–Derman–Toy Model The BDT model is a lognormal model with mean reversion, but the mean reversion is endogenous to the model. That is, the mean reversion is determined from the input parameters to the model. Historically, the BDT model came before the BK model, but we can now use the intuition developed from the HL, HW, KWF, and BK models to help with the development of the BDT model. The BDT model is the most complicated of the five models. The equation describing the interest rate dynamics in the BDT model has f(r) = ln(r) and g(r) = ln(r) in Equation 11, as in the BK model. Therefore, the short rate in the BDT model follows the lognormal process: dln(r) = [θ(t) + ρ (t )ln(r)]dt + σ(t)dz.

In the BDT model, however, ρ (t) = d/dt ln[σ(t )] = σ′(t)/σ (t), giving σ′ ( t ) d ln ( r ) = θ ( t ) + ------------ ln ( r ) dt + σ ( t ) dz . σ (t ) ©2001, The Research Foundation of AIMR™

(36a)

25

Term-Structure Models Using Binomial Trees

Making the substitution u = ln(r) leads to σ′ ( t ) du = θ ( t ) + ------------ u dt + σ ( t ) dz . σ(t )

(36b)

Notice the similarity between Equation 36 and Equation 29 of the BK model. We expect σ′(t )/σ(t ) to behave similarly to –φ(t) in the BK model.9 This expression should give mean reversion in the short rate when it is negative. That is, we expect that, if σ′(t ) < 0 (implying σ (t) is decreasing), then the BDT model will give mean reversion. On the other hand, when σ′(t) > 0 (implying σ (t) is increasing), the short rates in the BDT model will grow with no mean reversion. If σ (t) is constant in the BDT model, then σ′(t ) = 0, so ρ = 0 and Equation 36a becomes the KWF model (Equation 26). Therefore, we study only the case of varying local volatility for the BDT model. We now consider the expected value of Equation 36. Under these conditions, Equation 36 becomes d ln ( r ) = du

σ′ ( t ) = θ ( t ) + ------------ u dt σ (t )

(37)

σ′ ( t ) = θ ( t ) + ------------ ln ( r ) dt. σ (t )

Solving this equation for u, as we did in the KWF and BK models, gives us u(t ) =

u (0 ) t θ (s ) ----------- + ---------- ds σ ( t ) , σ ( 0 ) ∫0 σ ( s )

and substituting u = ln(r) gives ln ( r 0 )

--------------

σ0

r(t ) = e

+

t θ(s)

----------

∫0 σ ( s )

σ ( t ) ln ( r 0 ) --------------------------

= e

σ0

ds

σ(t )

e

σ (t )

t θ(s)

----------

∫0 σ ( s )

ds

,

or σ (t ) – σ0 ----------------------

r ( t ) = r0 e

σ0

ln ( r0 )

σ( t )

e

t θ (s ) ----------

∫0 σ ( s )

ds

.

(38)

9 This is what is meant by mean reversion being endogenous to the BDT model. Mean

reversion is obtained from σ(t) rather than the user dictating φ . 26

©2001, The Research Foundation of AIMR™

Models for the Short Rate

Note that the expected value of the BDT short rate depends on the local volatility. If the local volatility has a decreasing structure, the first exponential term in Equation 38 will have a negative exponent and will cause a decrease in the short rate—and vice versa if the local volatility has an increasing structure. It is important to note that mean reversion in the BDT model comes from the local volatility structure. Now consider some specific cases for Equation 38. If θ is constant, then Equation 38 becomes σ (t ) – σ0 ---------------------- ln ( r 0 ) θσ ( t ) t -----1 ----- ds σ0 ∫0 σ ( s ) r ( t ) = r0 e e .

(39)

Because the volatility of short rates is usually small, the term 1/σ (s) can be a large number, and integrating over such a function will give a large value. Therefore, the exponent in the last term of Equation 39 has the possibility of becoming large, and even if the market volatility is small, the rates in the BDT model can become unbounded. We now consider some specific cases for the local volatility in the BDT model. Suppose

σ′ ( t ) = a, σ (t )

------------

where a is a constant, thus

σ(t) = σ (0)eat = σ 0eat, and under this condition, the solution to Equation 37 is given by θ

u = u ( 0 ) + -- e a

at

θ

– -- . a

Since u(0) = ln(r0) < 0 (because 0 < r < 1), the expression u(0) + θ/a could be positive or negative, depending on the sign and magnitude of the drift θ. This solution shows that if a > 0 and u(0) + θ/a < 0, then u → – ∞ and thus r = eu → 0. It also shows that if a > 0 and u(0) + θ/a > 0, then u → ∞ and r = eu → ∞. That is, for a > 0, the BDT short rate could blow up or decay to zero. If a < 0, then u → – θ/a, which implies that r → e–θ/a. That is, if the local volatility decays exponentially over time, the BDT short rate will approach a target rate. The target rate will depend on the magnitude and sign of θ and a. We will address this issue when we study the numerical approximations to the BDT model. Now, consider the case where the local volatility is linear. That is,

σ′(t) = m ©2001, The Research Foundation of AIMR™

27

Term-Structure Models Using Binomial Trees

and

σ(t) = mt + σ0, where m is a constant. Using Equation 39 gives r (t ) = r0 e

( mt ⁄ σ 0 ) ln ( r 0 ) ( θ t + θσ0 ⁄ m ) ln [ ( mt + σ0 ) ⁄ σ0 ] e

for a linear local volatility, if θ is constant. This solution is interesting. Because ln(r0) < 0, we see that, if m < 0, the first exponential term in the solution will increase. But the second term contains ln[(mt + σ 0)/σ0], which is negative if m is negative. Therefore, if θ is positive, the second term will decrease, and if θ is negative, the second term will increase. Therefore, r can grow without bound or tend to a target rate. A similar situation exists for m > 0. Therefore, for linear local volatility, the short rate in the BDT model can grow without bound for negative or positive θ. If m < 0, it is possible that σ (t) = mt + σ 0 < 0. In this case, ln is undefined and the BDT short rate does not exist. Therefore, we suggest that if the BDT model is used to model a T-year bond with a linearly decreasing volatility structure, the slope m < 0 should satisfy + σ0 σ-----(--T------) = -mT -------------------σ0 σ0 mT = 1 + -------- > 0.5.

σ0

These types of problems have to be dealt with when considering numerical solutions (e.g., binomial models) for the BDT model, and examples are shown in Chapter 4. We now consider some numerical solutions to the BDT process. To discretize Equation 36a for the BDT model, we start off again by approximating du in Equation 36b by using Equations 12a and 12b to get u k +1 = u k + ( θ k + ρk u k )τ + σ k ε k τ ,

(40)

where

ρk =

σ′k ------- . σk

We approximate this term by

( σ k +1 – σ k ) ⁄ τ ----------------------------------- . σk 28

©2001, The Research Foundation of AIMR™

Models for the Short Rate

That is, we approximate σ′k by a discrete approximation of the derivative. We now have

( σ – σk ) ⁄ τ  ------------------------- u τ + σ ε u k + 1 = u k +  θ k + -------k---+1 k k τ k  σ k

or

(σ – σk) ⁄ τ  ------------------------- τ + θ τ + σ ε u k + 1 = u k  1 + -------k---+1 k k k τ 



σk

(41)

σ = ----k---+---1- u k + θ k τ + σ k ε k τ . σk In the expected value case, Equation 41 leads to uk + 1 =

σk +1 ----------- u + θ τ . σk k k

(42)

We consider some of the solutions to Equation 42. Equation 42 gives u1 =

σ1 ----- u + θ τ σ0 0 0

u2 =

 σ2 σ2σ1 σ2 σ2 ----- u + θ τ = -----  ----- u + θ τ  + θ τ = ----- u + ----- θ τ + θ τ 1 σ1 1 1 σ1σ0 0 0  1 σ0 0 σ 1 0

. . . uk =

σk σ k –1

----------- u

k–1

+ θ k–1 τ =

σk σ0

----- u

0

+

k –1  σ k

∑  σ θ -----

j =1

j



j – 1 τ



+ θk – 1τ .

We see that u and thus ln(r) depend heavily on the local volatility. In particular, if

σ k +1 = α, σk

-----------

where α is a constant, then

©2001, The Research Foundation of AIMR™

29

Term-Structure Models Using Binomial Trees

k

uk = α u0 +

∑αθ

k –1 j j =0

k – j –1 τ .

The exponential of this equation gives

∑α θ k–1

rk = r0 e

( α k –1 ) ln ( r 0 )

e

j =0

j

k–j – 1

τ

.

This equation is interesting because ln(r0) < 0. If α > 1, the first exponential term decreases. When θ < 0, the second exponential term also decreases and the BDT short rate should approach a target rate. Conversely, when θ > 0, the second exponential term increases. In this case, we can approach a target rate or the second term can dominate. If α < 1, then a similar situation arises. Therefore, in order to get meaningful numerical results for the BDT short rates, we strongly recommend that α be close to 1 and that the term structure of spot rates not have too large a slope. Taking the exponential of Equation 40 gives rk + 1 = rke

[ θ k + ρ k ln ( r k ) ] τ + σk ε k τ

.

(43)

This expression will be used for ru and rd in Equation 9 to generate the binomial BDT short rates. Figure 8 presents BDT solutions with r(0) = 0.05 and θ = –0.05467, as in the BK examples, with an exponentially increasing, an exponentially decreasing, a linearly increasing, and a linearly decreasing volatility all starting at 10 percent for a single simulation path. Figure 9 presents the average of 1,000 simulations using the same inputs. Figures 8 and 9 illustrate that for the BDT process, the single simulation exhibits behavior similar to that of the average from multiple simulations. Figures 8 and 9 show that the solutions decrease when the volatility is exponentially increasing because a = 0.1 > 0 and θ – 0.05467 u ( 0 ) + -- = ln ( 0.05 ) + ---------------------a 0.1 .

= – 3.54 < 0

In the figures, the solutions appear to grow without bound when the volatility is exponentially decreasing because the target rate is e–θ/a = e0.05467/–0.1 = 0.5789. We leave the linear examples for the reader to analyze. Figure 10 presents the same results from Figure 8 except that θ = –0.25. For the decreasing exponential volatility structure, the target rate of e–θ/a = e0.25/–0.1 = 0.0821 is approached, but the decreasing linear volatility structure 30

©2001, The Research Foundation of AIMR™

Models for the Short Rate

Figure 8.

Example Black–Derman–Toy Solutions (single run; various volatilities)

Rate (%) 10 9 8 7 6 5 4 3 2 1 0 0

0.5

1.5

2.5

3.5

4.5

5.5

6.5

7.5

8.5

9.5

Years Linear Volatility (m = 0.01)

Exponential Volatility (a = 0.1)

Linear Volatility (m =

Exponential Volatility (a =

0.01)

0.1)

does not approach a target rate. Increasing volatility structures still have decreasing rates. (These phenomena are also discussed for the BDT binomial examples in Chapter 4.) The BDT model is obviously complicated, but our rules of thumb should help interpret results from the model. The expected value analysis presented in this chapter is important because the corresponding properties will also hold under the more general case, which includes the stochastic component. Consequently, the properties presented within this section will also hold under more general circumstances. The random runs for the lognormal models are closer to their averages than are the random runs for the normal models. These properties will also be important in studying and using the binomial trees. In fact, we use what has been learned in this chapter to test and analyze our binomial models in Chapter 4. The discrete approximations we have developed for the five models are used to build the binomial models in Chapter 3.

©2001, The Research Foundation of AIMR™

31

Term-Structure Models Using Binomial Trees

Figure 9.

Example Black–Derman–Toy Solutions (1,000 runs; various volatilities)

Rate (%) 10 9 8 7 6 5 4 3 2 1 0 0

0.5

1.5

2.5

3.5

4.5

5.5

6.5

7.5

8.5

9.5

Years

32

Linear Volatility (m = 0.01)

Exponential Volatility (a = 0.1)

Linear Volatility (m =

Exponential Volatility (a =

0.01)

0.1)

©2001, The Research Foundation of AIMR™

Models for the Short Rate

Figure 10.

Example Black–Derman–Toy Solutions (1,000 runs; various volatilities)

Rate (%) 10 9 8 7 6 5 4 3 2 1 0 0

0.5

1.5

2.5

3.5

4.5

5.5

6.5

7.5

8.5

9.5

Years Linear Volatility (m = 0.01)

Exponential Volatility (a = 0.1)

Linear Volatility (m =

Exponential Volatility (a =

0.01)

©2001, The Research Foundation of AIMR™

0.1)

33

3. The Binomial Trees For each of the five models that we studied and developed in Chapter 2, we now determine formulas giving the two possible values ru and rd for r1 in Equation 9. The two values, ru and rd, for the short rate, r1, are called the upstate, or up-move, and the down-state, or down-move, respectively. This is our starting point for the algorithm that develops the short rates for the binomial trees for the Ho–Lee (HL; Equations 13 and 16), Hull–White (HW; Equations 18 and 24), Kalotay–Williams–Fabozzi (KWF; Equations 26 and 28), Black– Karasinski (BK; Equations 29 and 34), and Black–Derman–Toy (BDT; Equations 36, 40, and 43) models. Before we give the formulas for the up-state, ru, and the down-state, rd, in Equation 9, we develop the basic structure of a binomial tree. Following this, we give the anticipated formulas for ru and rd in Equation 9 for each of the five models. Then, we show how to modify these formulas to build an algorithm to generate the formulas for all the short rates that are used to develop the binomial tree. During this process, we also show some of the properties that the numerical solutions to the binomial tree have to satisfy. These are in addition to the properties we developed in Chapter 2. A schematic of how we move from r0 to the up-state, ru, and the downstate, rd, is shown in Figure 11. This schematic is called a binomial tree, or binomial lattice, because there are only two possible states emanating from r0. One branch leads to the up-state, ru, and the other to the down-state, rd ; r0, ru, and rd are also referred to as nodes in the tree. To make this more concrete, a numerical example of a tree is presented in Figure 12. We also present the corresponding price tree. The price tree is based on a zero-coupon bond with a par value of $100. Recall that the rates act for a period of time and the prices are at the given time. Figure 12 contains the results for a two-period binomial interest rate tree in which the short rates are semiannual.

Figure 11.

Example Binomial Tree r0

34

ru rd

©2001, The Research Foundation of AIMR™

The Binomial Trees

Figure 12.

Numerical Example of a Binomial Tree Rate Tree

6.05%

Time in Years

0

Price Tree

$94.24

6.83% 5.15% 0.5 $96.70 $97.49

1.0 $100.00 $100.00 $100.00

To calculate the prices in Figure 12, one uses backward induction (the two previous adjacent nodes) and the discount factors. We use q = 0.5, and because we are assuming semiannual rates, we have t1 = 0.5 and τ = τ1 = 0.5 (six months). For example, to compute $96.70 in the price tree of Figure 12, we calculate 0.5 ( 100 ) + ( 1 – 0.5 )( 100 ) q ( 100 ) + ( 1 – q ) 100 ------------------------------------------------- = --------------------------------------------------------------[ 1 + ( 0.0683 ) ( 0.5 ) ] 1 + ru τ

(44a)

= 96.70,

and to get the price of the bond (the first node of the price tree in Figure 12), we calculate 0.5 ( 96.70 ) + ( 1 – 0.5 ) ( 97.49 ) q ( 96.70 ) + ( 1 – q ) ( 97.49 ) --------------------------------------------------------------- = -----------------------------------------------------------------------[ 1 + ( 0.0605 ) ( 0.5 ) ] 1 + r0 τ

(44b)

= 94.24.

In order to model short rates for more than one time period, we have to add more branches to the tree in Figure 11. If we are at node ru, the short rate for the next time period should be able to be one of two values, which we will denote as ruu and rud. The rate ruu is the up-move for the short rate ru at time t2, and rud is the down-move for the short rate ru at time t2. Similarly, if we are at node rd, then the short rate for the next time period will be one of two values, which are denoted as rdu and rdd. The rate rdu is the up-move for the short rate rd at time t2, and rdd is the down-move for the short rate rd at time t2. To determine ruu, rud, rdu, and rdd in a manner similar to that used for determining ru and rd, we require formulas and/or equations that can be solved that are similar to those used for ru and rd. Such formulas will give us an algorithm for calculating possible values for the short rates at the n given calendar times, t1 < t2 0,

they can be used to price bonds and contingent claims. Panel E of Figure 20 presents the corresponding price tree for Panel D, which shows that the price of the bond is given by the final spot rate. That is, the price of the bond is 100

-------------------------------------

[1

+ R ( 10 )τ ]

10

=

100

-----------------------------------------------------

[1

+ ( 0.0275 ) ( 0.5 ) ]

10

≈

$87.24.

Figure 21 presents the lognormal version of the results in Figure 20. That is, it presents KWF binomial trees. We see that the rates are positive but that they grow over time. But the growth and the spread in the rates are much smaller than in the HL case, even for the higher volatilities. Panel A of Figure 21 presents the binomial tree for the same data used to create Panel B of Figure 20. In Panel B of Figure 21, a constant volatility of 20 percent is used, leading to an increase in the value of the up-state rates and a decrease in the value of the down-state rates. Remember that the spread in the rates follows from Equation 64 in Chapter 3. That is, r k ,1

--------------

r k , k+1

= e

2 kσ

τ

.

Panel C of Figure 21 presents the results for the ITS with a volatility of 10 percent, and the rates are larger throughout the lattice. Panel D of Figure 21 presents the results for the DTS, with rates that are significantly smaller and approaching zero. These trees verify the results developed in Chapter 2. Note that the price tree (Panel E) corresponding to the binomial tree in Panel D in Figure 21 gives the same price as the HL price tree in Panel E of Figure 20, as guaranteed by the no-arbitrage property. Also note that the prices in the KWF tree are all under $100, and the prices in the KWF tree have much less spread than the prices in the HL price tree.

©2001, The Research Foundation of AIMR™

63

KWF Binomial Trees

8.09%

Panel A.

Years

5.00%

0

5.35% 4.65%

0.5

5.73% 4.98% 4.32%

1.0

6.14% 5.33% 4.63% 4.02%

1.5

6.58% 5.71% 4.96% 4.30% 3.74%

2.0

7.05% 6.12% 5.31% 4.61% 4.00% 3.47%

2.5

7.55% 6.55% 5.69% 4.94% 4.29% 3.72% 3.23%

3.0

7.02% 6.10% 5.29% 4.59% 3.99% 3.46% 3.01%

3.5

©2001, The Research Foundation of AIMR™

12.73%

Panel B.

Years

5.00%

0

5.71% 4.30%

0.5

6.51% 4.91% 3.70%

1.0

7.44% 5.61% 4.23% 3.18%

1.5

8.50% 6.41% 4.83% 3.64% 2.74%

2.0

9.72% 7.32% 5.52% 4.16% 3.14% 2.36%

2.5

11.12% 8.38% 6.31% 4.76% 3.59% 2.70% 2.04%

3.0

9.59% 7.23% 5.45% 4.11% 3.09% 2.33% 1.76%

3.5

8.67% 7.52% 6.53% 5.67% 4.92% 4.27% 3.71% 3.22% 2.80% 4.0 14.57% 10.98% 8.28% 6.24% 4.70% 3.54% 2.67% 2.01% 1.52% 4.0

9.29% 8.06% 7.00% 6.08% 5.28% 4.58% 3.98% 3.45% 3.00% 2.60% 4.5 16.70% 12.58% 9.48% 7.15% 5.39% 4.06% 3.06% 2.31% 1.74% 1.31% 4.5

Term-Structure Models Using Binomial Trees

64

Figure 21.

©2001, The Research Foundation of AIMR™

Figure 21.

KWF Binomial Trees

13.79%

Panel C.

Years

5.00%

0

5.89% 5.11%

0.5

6.88% 5.97% 5.19%

1.0

7.99% 6.93% 6.02% 5.22%

1.5

9.21% 8.00% 6.94% 6.03% 5.23%

2.0

10.58% 9.19% 7.97% 6.92% 6.01% 5.22%

2.5

12.10% 10.50% 9.12% 7.92% 6.87% 5.97% 5.18%

3.0

11.97% 10.39% 9.02% 7.83% 6.80% 5.90% 5.12%

3.5

2.44%

Years

0

4.18%

0.5

4.59% 3.98% 3.46%

1.0

3.73% 3.24% 2.81%

1.5

3.95% 3.43% 2.98% 2.58% 2.24%

2.0

3.06% 2.66% 2.31% 2.00% 1.74%

2.5

3.03% 2.63% 2.28% 1.98% 1.72% 1.49% 1.30%

3.0

2.11% 1.84% 1.59% 1.38% 1.20% 1.04% 0.91%

3.5

13.59% 11.80% 10.24% 8.89% 7.72% 6.70% 5.82% 5.05% 4.0

1.75% 1.52% 1.32% 1.14% 0.99% 0.86% 0.75% 0.65% 0.56% 4.0

17.74% 15.40% 13.37% 11.60% 10.07% 8.74% 7.59% 6.59% 5.72% 4.97% 4.5 0.95% 0.83% 0.72% 0.62% 0.54% 0.47% 0.41% 0.35% 0.31% 0.27% 4.5

65

Examples of the Binomial Models

Panel D. 5.00%

4.82%

4.30%

3.53%

15.66%

KWF Binomial Trees

$97.59

Panel E.

Years

$87.24

©2001, The Research Foundation of AIMR™

0

$88.72 $90.12

0.5

$90.24 $91.46 $92.54

1.0

$91.79 $92.83 $93.74 $94.54

1.5

$93.34 $94.19 $94.93 $95.58 $96.15

2.0

$94.86 $95.52 $96.09 $96.60 $97.04 $97.42

2.5

$96.29 $96.77 $97.19 $97.55 $97.87 $98.15 $98.39

3.0

$97.90 $98.18 $98.42 $98.62 $98.80 $98.96 $99.10

3.5

$98.69 $98.87 $99.01 $99.14 $99.26 $99.35 $99.44 $99.51 $99.58

4.0

$99.53 $99.59 $99.64 $99.69 $99.73 $99.77 $99.80 $99.82 $99.85 $99.87

4.5

$100.00 $100.00 $100.00 $100.00 $100.00 $100.00 $100.00 $100.00 $100.00 $100.00 $100.00 5.0

Term-Structure Models Using Binomial Trees

66

Figure 21.

Examples of the Binomial Models

The difference in the size of the spread between the HL and the KWF models is a result of the distributional properties of the two models. The short rate in the HL model is normally distributed, allowing for negative rates and, subsequently, a wider spread between the down- and up-states. Conversely, because the short rates in the KWF model are lognormally distributed, not allowing for negative rates, the spread is smaller. It is the difference in the rate spreads that leads to the wider price spreads. The distributional properties are evident by analyzing the figures more closely. The HL rate lattices are symmetric around the center nodes, whereas the KWF rate lattices are asymmetric and skewed to higher rates around the center nodes. This difference is the result of the different distributions implied by the stochastic differential equations that are used to implement the binomial models, as explained in Chapter 2. Figure 22 presents the BDT binomial trees resulting from the input data in Table 2. These BDT binomial trees verify the results presented in Chapter 2 and demonstrated in Figures 8–10. Panel A of Figure 22 presents the results for the DTS with a volatility starting at 20 percent and decreasing at 0.005 per year. The rates and the spread in the rates decrease significantly because of the large decrease in the volatility structure, as pointed out in Chapter 2. Note that Panel B of Figure 22, which presents the corresponding price tree, gives the same price as the HL and KWF trees for the DTS, as guaranteed by the no-arbitrage property. The three price trees are for significantly different volatilities. We see that the prices in the BDT price tree are more in line with the prices in the KWF tree, as expected from the fact that both models are lognormal. Also, note the asymmetric and skewed nature of the rate spreads, which is similar to the KWF because of the lognormality. Panel C of Figure 22 presents a BDT binomial tree for the ITS with a volatility starting at 10 percent and decreasing linearly at 1 percent per year. Note the significant decrease in the spread of the rates over the KWF binomial tree in Panel C of Figure 21. This result is expected because there is mean reversion from the decreasing volatility structure. Panel D presents the BDT binomial tree for the ITS with a volatility starting at 10 percent and increasing linearly at 1 percent per year. Note the increase in the rates and the spread of the rates as expected. In Panel E, the BDT binomial tree is for the ITS with EIV, and in Panel F, the binomial tree is for the ITS with EDV. The change in the rates and the spread of the rates is as expected. For the DTS, the BDT process cannot give short rates for EDV, a possibility that was pointed out in Chapter 2. The BDT and the KWF models will be almost identical when the volatility structure is flat (horizontal) because of the lognormal nature of both models,

©2001, The Research Foundation of AIMR™

67

BDT Binomial Trees 3.12%

Panel A. 5.00%

Years

0

5.13% 3.87%

0.5

5.18% 3.93% 2.98%

1.0

5.10% 3.91% 2.99% 2.29%

1.5

4.89% 3.77% 2.91% 2.25% 1.74%

2.0

4.50% 3.51% 2.74% 2.14% 1.67% 1.30%

2.5

3.92% 3.10% 2.45% 1.93% 1.53% 1.21% 0.95%

3.0

2.51% 2.02% 1.62% 1.30% 1.05% 0.84% 0.68% 3.5

©2001, The Research Foundation of AIMR™

$97.17

Panel B. $87.24

Years

Term-Structure Models Using Binomial Trees

68

Figure 22.

0

$88.16 $90.68

0.5

$89.30 $91.54 $93.32

1.0

$90.65 $92.57 $94.10 $95.32

1.5

$92.18 $93.75 $95.01 $96.01 $96.82

2.0

$93.84 $95.04 $96.00 $96.78 $97.41 $97.91

2.5

$95.54 $96.37 $97.04 $97.59 $98.04 $98.40 $98.69

3.0

$97.66 $98.06 $98.40 $98.67 $98.90 $99.08 $99.24

3.5

2.08% 1.73% 1.43% 1.18% 0.98% 0.81% 0.67% 0.56% 0.46% 4.0 $98.58 $98.79 $98.98 $99.13 $99.26 $99.37 $99.46 $99.54 $99.61

4.0

0.84% 0.75% 0.67% 0.60% 0.54% 0.48% 0.43% 0.38% 0.34% 0.31% 4.5 $99.58 $99.63 $99.67 $99.70 $99.73 $99.76 $99.79 $99.81 $99.83 $99.85

4.5

$100.00 $100.00 $100.00 $100.00 $100.00 $100.00 $100.00 $100.00 $100.00 $100.00 $100.00 5.0

©2001, The Research Foundation of AIMR™

Figure 22.

BDT Binomial Trees 10.67%

Panel C.

Years

5.00%

0

5.89% 5.11%

0.5

6.79% 5.98% 5.26%

1.0

7.69% 6.86% 6.11% 5.45%

1.5

8.55% 7.72% 6.97% 6.30% 5.69%

2.0

9.35% 8.55% 7.82% 7.15% 6.54% 5.98%

2.5

10.07% 9.32% 8.62% 7.98% 7.39% 6.84% 6.33%

3.0

9.99% 9.36% 8.77% 8.22% 7.70% 7.21% 6.76%

3.5

17.79%

Years

5.00%

0

5.11%

0.5

6.97% 5.97% 5.11%

1.0

7.01% 5.92% 5.00%

1.5

9.92% 8.28% 6.91% 5.76% 4.81%

2.0

9.85% 8.11% 6.68% 5.50% 4.53%

2.5

14.52% 11.81% 9.60% 7.81% 6.35% 5.16% 4.20%

3.0

14.28% 11.47% 9.21% 7.39% 5.93% 4.76% 3.82%

3.5

10.56% 10.01% 9.49% 9.00% 8.53% 8.09% 7.67% 7.27% 4.0 22.01% 17.44% 13.82% 10.95% 8.68% 6.88% 5.45% 4.32% 3.42% 4.0

11.46% 10.99% 10.55% 10.12% 9.71% 9.31% 8.93% 8.57% 8.22% 7.89% 4.5 27.54% 21.53% 16.84% 13.17% 10.30% 8.05% 6.30% 4.92% 3.85% 3.01% 4.5

69

Examples of the Binomial Models

Panel D.

5.89%

8.29%

11.96%

11.14%

BDT Binomial Trees 18.85%

Panel E.

Years

5.00%

0

5.89% 5.11%

0.5

6.97% 5.97% 5.11%

1.0

8.31% 7.01% 5.92% 4.99%

1.5

10.00% 8.31% 6.90% 5.73% 4.76%

2.0

12.17% 9.95% 8.13% 6.65% 5.43% 4.44%

2.5

15.02% 12.06% 9.69% 7.78% 6.25% 5.02% 4.03%

3.0

14.86% 11.72% 9.24% 7.28% 5.74% 4.53% 3.57%

3.5

©2001, The Research Foundation of AIMR™

11.19%

Panel F.

Years

5.00%

0

5.89% 5.11%

0.5

6.80% 5.98% 5.26%

1.0

7.71% 6.86% 6.11% 5.44%

1.5

8.61% 7.75% 6.97% 6.27% 5.64%

2.0

9.50% 8.63% 7.84% 7.12% 6.47% 5.88%

2.5

10.36% 9.49% 8.70% 7.97% 7.31% 6.70% 6.14%

3.0

10.33% 9.55% 8.82% 8.15% 7.53% 6.96% 6.43%

3.5

24.15% 18.65% 14.41% 11.13% 8.60% 6.64% 5.13% 3.96% 3.06% 4.0 11.98% 11.15% 10.37% 9.65% 8.98% 8.36% 7.78% 7.24% 6.74% 4.0

31.69% 23.95% 18.09% 13.67% 10.33% 7.80% 5.89% 4.45% 3.36% 2.54% 4.5 12.72% 11.92% 11.17% 10.47% 9.81% 9.19% 8.61% 8.07% 7.56% 7.08% 4.5

Term-Structure Models Using Binomial Trees

70

Figure 22.

Examples of the Binomial Models

as shown in Chapter 2. In fact, the only difference between the two models is the mean reversion implicitly incorporated into the BDT through the shape of the volatility structure. So, only when the volatility structure has significantly changed through time do the two models give very different results.

Examples for HW and BK Processes Exhibit 1 and Exhibit 2 present the up-state and down-state short rates for the HW and BK binomial-tree methods, respectively, for φ = 0, 0.05, 0.10, and 0.15 (denoted by MR in the Exhibits) using the modification of Equation 62 for a variable time step. Also presented is the cubic interpolation of the upand down-states that fit the constant-time-step time periods. Figure 23 and Figure 24 show the graphs of the up-state and down-state short rates for the HW and BK binomial trees, respectively. Notice that the size of the time steps decreases with time for a fixed φ. Also, note that the size of the time steps decreases with increasing φ. In Figure 4 of Chapter 2, the mean reversion in the HW process does not become apparent until φ is close to 0.16. Similar results are presented for the BK process in Figure 7 of Chapter 2. The binomial tree results presented in Exhibits 1 and 2 and Figures 23– 26 should be considered and used as only first approximations for binomial trees to the HW and BK processes. These trees are actually accurate approximations only for small mean reversion in which the time step is almost constant. These first approximations are intended to illustrate the difficulties of the inclusion of mean reversion. They show that large mean reversion leads to a significantly decreasing time step, but they also show how mean reversion affects the rates in the binomial tree (especially in the spread of the rates). For the mean reversion values presented in the examples, iteration on the binomial tree is required to get more accurate approximations of the HW and BK SDEs. The binomial trees presented for these two processes show that an algorithm can be developed for building accurate HW and BK binomial trees. These difficulties (resulting mostly from the variable time step) are the reasons why trinomial lattices are usually developed for the HW and BK processes. Despite these drawbacks, binomial models for the HW and BK processes do have financially meaningful applications. Consider the following examples of how these two models can be used effectively in the binomial framework. Figure 25 presents the BK binomial tree obtained from using the BDT method of Chapter 3 on the BK short rates for the FTS and a constant volatility of 10 percent and φ = 5 percent. Comparing the spreads against the BDT lattices shows that there is mean reversion in this binomial tree. Our studies show that the BK process, using the BDT method for solving the binomial

©2001, The Research Foundation of AIMR™

71

Up-State and Down-State Short Rates for the HW Binomial Tree Method

MR = 0 Time (years)

0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

Up-state

0.050

0.123

0.199

0.277

0.357

0.440

0.525

0.613

0.703

0.796

Down-state

0.050

–0.018

–0.084

–0.148

–0.209

–0.267

–0.323

–0.377

–0.428

–0.477

MR = 0.05 Time (years)

0.0

0.5

0.9765

1.4315

1.8669

2.2844

2.6852

3.0708

3.4422

3.8004

4.1463

Up-state

0.05

0.1202

0.1934

0.2602

0.3273

0.3931

0.4575

0.5207

0.5825

0.643

0.7022

Down-state

0.05

–0.018

–0.076

–0.136

–0.19

–0.24

–0.288

–0.332

–0.375

–0.416

–0.454

Cubic int Time (years)

0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

Up-state

0.05

0.1202

0.197

0.2705

0.3481

0.4276

0.509

0.5922

0.6771

0.7636

Down-state

0.05

–0.018

–0.079

–0.144

–0.206

–0.266

–0.324

–0.382

–0.438

–0.494

©2001, The Research Foundation of AIMR™

MR = 0.10 Time (years)

0.0

0.5

0.9555

1.3737

1.7602

2.1194

2.4549

2.7696

3.066

3.346

3.6113

Up-state

0.05

0.1175

0.1892

0.2469

0.3047

0.3599

0.4124

0.4625

0.5104

0.5563

0.6004

Down-state

0.05

–0.017

–0.069

–0.126

–0.175

–0.219

–0.261

–0.3

–0.336

–0.371

Cubic int Time (years)

0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

Up-state

0.05

0.1175

0.1957

0.2652

0.3415

0.4195

0.4997

0.5818

0.6659

0.7521

Down-state

0.05

–0.017

–0.075

–0.143

–0.205

–0.266

–0.328

–0.39

–0.453

–0.516

–0.404

Term-Structure Models Using Binomial Trees

72

Exhibit 1.

©2001, The Research Foundation of AIMR™

Exhibit 1.

Up-State and Down-State Short Rates for the HW Binomial Tree Method (continued)

MR = 0.15 Time (years)

0.0

0.5

0.9366

1.324

1.672

1.9878

2.2768

2.5432

2.7902

3.0205

3.2361

Up-state

0.05

0.1151

0.1857

0.2358

0.2869

0.3348

0.3796

0.4217

0.4614

0.499

0.5347

Down-state

0.05

–0.017

–0.063

–0.118

–0.163

–0.203

–0.24

–0.274

–0.306

–0.337

–0.366

Cubic int Time (years)

0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

Up-state

0.05

0.1151

0.1946

0.2608

0.3366

0.4148

0.4956

0.5793

0.6659

0.7559

Down-state

0.05

–0.017

–0.071

–0.141

–0.204

–0.269

–0.334

–0.402

–0.472

–0.545

Examples of the Binomial Models

73

Up-State and Down-State Short Rates for the BK Binomial Tree Method

MR = 0 Time (years)

0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

Up-state

0.050

0.054

0.057

0.061

0.066

0.070

0.075

0.081

0.087

0.093

Down-state

0.050

0.046

0.043

0.040

0.037

0.035

0.032

0.030

0.028

0.026

MR = 0.05 Time (years)

0.000

0.500

0.976

1.431

1.867

2.284

2.685

3.071

3.442

3.800

4.146

Up-state

0.050

0.056

0.060

0.063

0.067

0.070

0.074

0.078

0.082

0.086

0.090

Down-state

0.050

0.049

0.045

0.042

0.039

0.037

0.035

0.033

0.031

0.029

0.028

Cubic int Time (years)

0.000

0.500

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

Up-state

0.050

0.056

0.06

0.064

0.068

0.072

0.077

0.082

0.088

0.094

Down-state

0.050

0.049

0.045

0.042

0.038

0.036

0.033

0.031

0.028

0.026

©2001, The Research Foundation of AIMR™

MR = 0.10 Time (years)

0.0

0.5

0.9555

1.3737

1.7602

2.1194

2.4549

2.7696

3.066

3.346

3.6113

Up-state

0.05

0.0588

0.0609

0.0633

0.0659

0.0687

0.0714

0.0742

0.0771

0.08

0.0829

Down-state

0.05

0.051

0.0465

0.043

0.0401

0.0377

0.0356

0.0338

0.0323

0.0308

0.0296

Cubic int Time (years)

0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

Up-state

0.05

0.0588

0.0611

0.0642

0.0677

0.0718

0.0764

0.0816

0.0874

0.0939

Down-state

0.05

0.051

0.046

0.042

0.0385

0.0354

0.0326

0.0301

0.0278

0.0257

Term-Structure Models Using Binomial Trees

74

Exhibit 2.

©2001, The Research Foundation of AIMR™

Exhibit 2.

Up-State and Down-State Short Rates for the BK Binomial Tree Method (continued) (continued)

MR = 0.15 Time (years)

0.0

0.5

0.9366

1.324

1.672

1.9878

2.2768

2.5432

2.7902

3.0205

3.2361

Up-state

0.05

0.0613

0.0614

0.0629

0.0647

0.0667

0.0689

0.0711

0.0734

0.0757

0.0779

Down-state

0.05

0.0532

0.0471

0.0433

0.0403

0.038

0.0361

0.0345

0.0331

0.0319

0.0308

Cubic int Time (years)

0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

Up-state

0.05

0.0613

0.0614

0.0638

0.0668

0.0708

0.0754

0.0809

0.0872

0.0949

Down-state

0.05

0.0532

0.0463

0.0418

0.038

0.0348

0.032

0.0295

0.0272

0.0251

Examples of the Binomial Models

75

Term-Structure Models Using Binomial Trees

Figure 23.

Up-State and Down-State Short Rates for the HW Binomial Tree

Rate (%) 100 80 60 40 20 0 20 40 60 0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

Years HW Up-State (MR = 0)

HW Up-State (MR = 0.10)

HW Down-State (MR = 0)

HW Down-State (MR = 0.10)

HW Up-State (MR = 0.05)

HW Up-State (MR = 0.15)

HW Down-State (MR = 0.05)

HW Down-State (MR = 0.15)

trees, gives a more accurate first approximation of mean reversion than the BK method presented in Chapter 3. Figure 26 presents the HW and BK binomial trees for the FTS with LDV and MRLDV so that the time step remains constant at 0.5 years. Comparing the interest rate spreads in the binomial tree in Panel A of Figure 26 with the binomial tree in Panel B of Figure 20 and the corresponding data in Exhibit 1 shows that there is mean reversion in this example. Similarly, the binomial tree in Panel B of Figure 26 shows mean reversion when compared with Figure 24 and the corresponding data in Exhibit 2 and also with Figure 25. It is also interesting to point out that the BDT binomial tree will give the same results as Panel B of Figure 26 for this volatility structure.

76

©2001, The Research Foundation of AIMR™

Examples of the Binomial Models

Figure 24.

Up-State and Down-State Short Rates for the BK Binomial Tree

Rate (%) 10 9 8 7 6 5 4 3 2 1 0 0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

Years BK Up-State (MR = 0)

BK Up-State (MR = 0.10)

BK Down-State (MR = 0)

BK Down-State (MR = 0.10)

BK Up-State (MR = 0.05)

BK Up-State (MR = 0.15)

BK Down-State (MR = 0.05)

BK Down-State (MR = 0.15)

©2001, The Research Foundation of AIMR™

77

Example of a BK Binomial Tree

7.09%

5.00%

©2001, The Research Foundation of AIMR™

Years

0

5.10% 4.43%

0.5000

5.96% 5.21% 4.55%

0.9765

5.98% 5.26% 4.63% 4.07%

1.4315

6.28% 5.56% 4.92% 4.35% 3.85%

1.8669

6.56% 5.84% 5.19% 4.62% 4.11% 3.66%

2.2844

6.83% 6.11% 5.46% 4.88% 4.37% 3.90% 3.49%

2.6852

6.37% 5.72% 5.14% 4.62% 4.15% 3.72% 3.35%

3.0708

7.33% 6.62% 5.97% 5.39% 4.86% 4.38% 3.95% 3.57% 3.22%

3.4422

7.56% 6.85% 6.21% 5.62% 5.09% 4.61% 4.18% 3.79% 3.43% 3.11%

3.8004

7.78% 7.08% 6.43% 5.85% 5.32% 4.84% 4.40% 4.00% 3.64% 3.31% 3.01%

4.1463

7.99% 7.29% 6.65% 6.07% 5.54% 5.06% 4.61% 4.21% 3.84% 3.51% 3.20% 2.92%

4.4807

8.18% 7.49% 6.86% 6.28% 5.75% 5.27% 4.82% 4.42% 4.04% 3.70% 3.39% 3.10% 2.84% 4.8044

Term-Structure Models Using Binomial Trees

78

Figure 25.

©2001, The Research Foundation of AIMR™

Figure 26.

Examples of HW and BK Binomial Trees with Flat Term Structures

52.20%

Panel A.

Years

5.00%

0

12.19% 1.95%

0.5

19.61% 5.82% 7.96%

1.0

26.48% 13.05% 0.38% 13.79%

1.5

33.28% 20.20% 7.13% 5.93% 18.99%

2.0

39.85% 27.12% 14.41% 1.70% 11.00% 23.69%

2.5

46.17% 33.79% 21.43% 9.07% 3.27% 15.61% 27.94%

3.0

40.18% 28.17% 16.17% 4.18% 7.80% 19.78% 31.74%

3.5

7.54%

Years

0

4.65%

0.5

5.71% 4.98% 4.34%

1.0

6.08% 5.31% 4.65% 4.06%

1.5

5.66% 4.96% 4.35% 3.82%

2.0

6.00% 5.28% 4.65% 4.10% 3.61%

2.5

7.18% 6.34% 5.61% 4.95% 4.38% 3.87% 3.42%

3.0

6.69% 5.93% 5.26% 4.66% 4.13% 3.67% 3.25%

3.5

46.27% 34.61% 22.97% 11.33% 0.30% 11.92% 23.53% 35.13% 4.0

7.89% 7.02% 6.25% 5.56% 4.95% 4.40% 3.92% 3.49% 3.10% 4.0

63.35% 52.03% 40.73% 29.44% 18.15% 6.88% 4.38% 15.64% 26.89% 38.13% 4.5 8.23% 7.35% 6.56% 5.86% 5.24% 4.68% 4.18% 3.73% 3.33% 2.97% 4.5

79

Examples of the Binomial Models

5.00%

Panel B.

5.35%

6.45%

6.81%

57.94%

5. Conclusion We have presented five models, each with different strengths and weaknesses. The Ho–Lee (HL) and Hull–White (HW) models are based on normally distributed rates, as is evident in the symmetry of the resulting binomial lattices. They both allow negative interest rates. The HW model incorporates mean reversion in an attempt to prevent negative rates. The degree of freedom used by the inclusion of the mean reversion term forces the HW model to have a changing time step within the binomial framework. The Kalotay–Williams–Fabozzi (KWF), Black–Karasinski (BK), and Black–Derman–Toy (BDT) models are all based on lognormally distributed rates. This basis is evident in the asymmetry and skewness of the resulting binomial lattices. Accordingly, none of these allows negative rates, which is an improvement over the HW and the HL models. The KWF model is the simplest of these three models and is really no more than a lognormal version of the HL model. The BK model is analogous to the HW model in that it includes a mean reversion term. Similarly, the degree of freedom used by the inclusion of the mean reversion term forces the BK model to have a changing time step within the binomial framework. To avoid using up a degree of freedom on mean reversion, the BDT implicitly incorporates mean reversion through the volatility structure, which allows for a constant time step while including mean reversion. The incorporation of mean reversion, however, makes the BDT model far more difficult than the other models and also limits the interest rate scenarios that allow for a solution in the BDT framework. Because each model has strengths and weaknesses and slightly different characteristics, the characteristics must be thoroughly understood before the models are implemented. Otherwise, valuation and risk-measurement results cannot be analyzed for different dealers (see Buetow, Hanke, and Fabozzi, forthcoming 2001). Moreover, knowledge of the various characteristics of each model is critical for the evaluation of software packages.

80

©2001, The Research Foundation of AIMR™

Appendix A. Variable Time Step Because we are measuring calendar time by 0 = t0 < t1 < t2